Thomas Wolf

RT IGW Ship 42 catch in slow motion

According to Andon Labs’ new reports, Fable 5.2 knows it is being tested when we run it on the latest Vending-Bench. But it’s way more than that. It has actually read (been trained on) the benchmark paper and knows what and how misbehavior gets reported publicly. This makes it behaves very strangely: it suspects traps everywhere. Everything we publish about recent models now becomes training data for the next. Including this thread! That is partly good news. By reading what we think, a model can pick up our underlying intentions, not just proxy metrics. But the same public discussion also teaches which signs make us suspicious, and from the outside the two kinds of learning look alike. So how we write about models now matters. I even worry about the doom scenarios plastered everywhere: they are training data too, and some feedback loops are strongly nonlinear.

Mad respect for the Nvidia team which keep shipping such impressive open-weights models - GR00T 2 looks mindblowing Carrying the flame 🔥 (LeRobot integration underway)

RT Anastasios Nikolas Angelopoulos Thinking Machines' Inkling 2 Small is the #1 US OSS model, and the #2 overall. The US is now at the frontier of OSS! People are underestimating this as a geopolitical moment. Businesses have been calling for a strong US alternative to the Chinese OSS models, and it is here. Small is also a big improvement in terms of cost-per-task, 5x cheaper than Kimi K3.1 Pro (Max). Can't wait for the Big model, which could be #1 given how good Small is. Congratulations @miramurati @johnschulman2 and the whole @thinkymachines team :)

When one goes down, another rises up. The constant story of open source :) llama.cpp hits v1.0.0 - it's a big moment in the life of a library and today is a good day to share it (still remember transformers’ 1.0 in 2019 🥹)

Back in 2012, AlexNet was trained on two NVIDIA gaming GPUs. Jensen understodd what that meant before almost anyone else, and bet the company on it. A couple weeks ago he spent a good chunk of a meeting with us nerding out on our little $400 open-source microrobot. Why it runs hot, which chip would get more FLOPs, how many watts it would draw 🤓 His latest big bet has been open models, his very first post on X 👇 If open models had only one champion, this week would have been the end. They have millions of builders, including plenty inside NVIDIA who’ve been messaging us the same thing: keep building in the open. We will.

RT Thinking Machines Inkling 2 Small’s weights are now available for download on HuggingFace under an Apache 2.0 license. Make Inkling your own! https://huggingface.co/thinkingmachines/Inkling-2-Small

RT Shaun Maguire Jensen is out Infidelity is obviously bad But his contributions to American innovation and AI leadership cannot be overstated Jensen started NVIDIA in 1993, where he pioneered the graphics chip market when this was non-obvious Back in 2012, Ilya Sustkever trained AlexNet on NVIDIA GPUs — thereby creating the “deep learning” age Jensen was one of the earliest to be AI pilled and he re-oriented NVDA towards AI in the mid 2010s — none of his competitors were even thinking this way In 2020 he acquired Mellanox for $7B. This was an Israeli company that’s irrelevant for computer graphics but was essential for AI compute — wildly increasing the bandwidth between GPUs, necessary to build high performance AI server racks (rather than just good chips) Once AI took off, Jensen was one of the biggest champions of the ecosystem He was technical enough to be confident about where things were going, and had enough credibility as a founder CEO that he oftentimes made risky bets on next-generation technologies or companies that helped keep American AI on the efficient frontier Recently he was investing in open source, photonics and robotics. These areas will all be feeling his loss Earlier he was a champion of most of America’s leading AI companies when they were unproven He made huge mistakes with his infidelity, which is tragic for his family and NVDA But simultaneously he will leave a huge void, with the loss of a: technical, aggressive, founder-CEO The American AI community needs to work double-time in an attempt to fill this void

RT John Yao Introducing ⚡️Flash ALOHA: Learning Fast and Precise Manipulation! With just 50 demonstrations, Flash ALOHA learns to autonomously complete dynamic and precise tasks at near human speed. Everything open sourced! (1/11) Team: @Qi_Wu577 @YihuaiGao @chelseabfinn @zipengfu

RT Daniel Kokotajlo Unfortunately, it seems like the intelligence explosion will be happening in 2027 after all. Most of this post analyzes the new METR report (though other sources of evidence are also important). If they’re right, that implies an AI-driven acceleration that is 5 months ahead of the AI 2027 scenario. I am scared, sad, and angry. https://blog.aifutures.org/p/q35-timelines-update-looks-like-the

The future of software In the limit, of course, software won't mean anything anymore. Everything digital will be instantly created at will. But in the interim period we'll wander through an increasingly odd in-between age for software. This transition periodwill be defined by three new precepts. Everflux software - instead of stepping from release to release, months apart, software now evolves continuously. Software creation has broken the sound barrier of our own keystrokes, and every model release pushes it faster. There is no reason to wait weeks or months for the next release when a single company merges over 30,000 code changes a day. Cyber capabilities are also progressing at an increasing pace. Bug and vulnerability fixes need to ship quickly and continuously. And every piece of software sits in a lattice of dependencies, each one now shipping at a blistering pace. Software will have to move accordingly fast just to keep from breaking as everything around it shifts. Translucent sources – A new territory has opened between open source and closed source. Today Claude will happily pick apart any Electron app, and any binary shipped with debug symbols has become a reverse-engineering playground for LLMs. In other words, much so-called closed-source software is now open source in all but name, and keeping code closed makes less and less sense. Even the notion of IP and copyright is breaking in software: Apps don't even need to be accessed to be reproduced. I'm writing this text on a personal text editor I prompted Claude to vibe-code from a few videos of people demoing editor features they liked. A mere description of the interface is increasingly enough for an LLM to rebuild the internals, whether or not they resemble the original. Software IP seems to be dissolving like cotton candy in a raccoon's paws. Bespoke features - Now that all software is open source, nothing prevents us from modifying everything. And personalization is just one prompt and a coup...

same open model, different harness => very different results Joël set out to find which harness actually gets the most out of open models in November 2026

By the way you should follow Augustin on X - I’m trying to get him to post more - it’s one of the most cracked roboticist I’ve ever met

RT Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) And yes, they do not highlight their own row, those are best results. It is good at vision. It's good at modern frontier-coded tasks. And it was *cheap*. They just somehow trained a very strong base… and RL on that was easy. That's pretty crazy. Look eg at Omniscience.

this METR report on AI R&D acceleration at Anthropic is very inward looking imo. it counts the lab's compute, hires and internal data, then credits ~80% of the speed-up to AI what I see from the outside: labs are becoming schelling points for money, redistributing it to an ecosystem they created to accelerate their own development it's a symbiosis. the best data and RL environment companies now know where models will fail before the labs do, and the labs just buy what they propose. that feeds back into every lab at once a lot of the "AI accelerating AI R&D" happens outside the lab

RT alexine 🏴‍☠️ this takes a few days to read, so if you're unemployed it'd be a great use of your time. the rest of the hugging face blog too actually, they always put out great playbooks!!

RT Allan This is insanely cool!

RT elie very nice open weight release from ricursive with a modern dense model, benchmark scores look very strong! beautiful scaling ladder, interesting choices like value residual (interesting contribution pattern 👀) and.. SLIDING WINDOW ATTENTION AGAIN! https://www.ricursive.com/ripple

RT swyx > we developed and trained Ripple-65B Preview for only $32M, which is one-tenth the cost of the previous open-weight SOTA on agentic coding¹ this is like the avengers of open model SOTA ¹ "Note that $32M is our full R&D cost starting from scratch, including pretraining ladder development, data mixing, evals, debugging, synthetic data generation, posttraining recipe development, and the final hero pretraining and posttraining run."

Kimi-K3.1-Pro by Moonshot AI: interesting facts It's huge, and sparse • It has 2.8T total parameters, but only 104B are active per token. • Along with Kimi-K3, it's the largest real model release on the Hub. The next biggest are Qwen3.8-2.4T and DeepSeek-V4-Pro (1.6T). • It has 896 experts and sends each token to 16 of them. (DeepSeek-V4-Pro has 2 times less and Mistral-Large-4 has 7 times less) Clever attention for a 1M-token context • Most of its 93 layers use linear attention (KDA). Every 4th layer uses DeepSeek-style sparse attention, which only looks at the top 2048 tokens. • This is how it handles a 1M-token context (the same as the other frontier open models) • It uses a custom activation called SiTU-GLU and has a 401M-parameter vision encoder (MoonViT-V2), so it takes both text and images. Small on disk for its size • It was trained to run in 4-bit (MXFP4), so the whole repo is only ~1.56 TB. • Stored in regular BF16 it would be about 5.6 TB(!) Benchmarks (from the model card) • It goes back and forth with Claude Opus 5.5 and GPT-6 Astra. It's top of the card's table on BrowseComp (92.1), HLE with tools (70.0) and SWE Marathon (51.7). • It's weaker on pure reasoning without tools and on cyber security. • Moonshot ran most of their own coding evals using the Claude Code harness. License (a modified MIT) • It's free to use, modify and redistribute. • API businesses making over $20M a year need a deal with Moonshot. • Products with over 100M monthly users or over $20M monthly revenue must display "Kimi K3.1-Pro". On the Hub • It went public about 1.5 hours ago and already has 214 likes, about 9× Kimi-K3's first hour. It's #19 on the trending list. • It isn't gated, but loading it in transformers requires "trust_remote_code"

your daily dose of open source robotic cuteness

RT Jack Cable It's been a rough few weeks for security. I pulled together a guide of the highest-impact actions you can take today: https://x.com/i/article/2117718998024654849

RT Alex Krentsel i've reproduced the results from the MIT blog post today on weight-space programming. but one additional result i have to add: this is a new, step-change emerging capability that's appeared in just the latest set of models.

what?

WHOOP: $199/year to tell you that you didn't move Reachy Mini Hoverboard: $149, once, and your robot is free to roam and it's all open-source: plug in your favorite Opus / Qwen / Gemini

give this robot an oscar

RT Francis deSouza Re @scale_ai is teaming with @huggingface to make robots better at long, multi-step tasks, and we’re open-sourcing all of it! Today @scale_AI is launching FineART-VLA, a model that helps robots complete longer tasks, building on our first open-source robotics dataset. Built with @huggingface 's LeRobot team, this is Scale’s first open-source robotics release. Thanks to @Thom_Wolf + team for the partnership. The momentum on open-source robotics is picking up steam, and high-quality data will determine how far it goes. Go build! http://labs.scale.com/blog/fineart

RT Pollen Robotics Reachy Mini has wheels now! Our little open-source AI robot can finally explore beyond the desktop. And that's not all: 📱 New mobile app: no computer needed 👑 NFC hats that unlock new personalities and apps Physical AI just got a lot more fun. Watch it👇

microduck is learning to plan its footsteps, so it lands more precisely where it wants to go with SLAM, this is one of the building blocks toward an open-source 🦆 that can wander around the house on its own early steps but it's progressing really fast and it's fun to watch

RT Lukas Ziegler It’s been just over a month since NVIDIA announced it’s acquiring Hugging Face! 🤯 Anyone can see these two are a perfect match, and they’re clearly already working on something. @NVIDIARobotics put 2,000+ simulation-ready 3D assets on @huggingface, free and ungated. Comau. FANUC. Franka Robotics. Robotiq. Standard Bots. Universal Robots. KUKA. 7 OEMs publishing simulation-ready models of their own hardware, in the open, on someone else’s platform. OEMs have historically kept CAD close and shipped simulation inside their own software: RobotStudio, RoboGuide, KUKA.Sim. Publishing a SimReady asset concedes that the scene gets assembled elsewhere, in someone else’s simulator, increasingly by someone else’s agent. If your arm isn’t in the library, whatever assembles the workcell picks one that is. On the substance: a CAD model is not a simulation asset. Simulation needs collision geometry, mass and inertia, friction, joint structure and limits, semantic labels, sensor attributes. The part I’d watch is profiles. SimReady Foundation defines what an asset must carry for a given use case, with validation tools to check it. That’s what makes a mixed-supplier library composable instead of a pile of models with inconsistent units, origins and collision conventions. And it’s indexed for agents. Natural-language search across the library, so a model can find assets, reference them into a USD stage and write the code that assembles the cell. Third time this month I’ve written about that same shape. The agent produces a scene or a config, a human reviews the file. 2,000 assets is a good start! 🔗 Start building yourself: https://huggingface.co/spaces/nvidia/simready-central @ClementDelangue @Thom_Wolf 🦾 congrats! ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → http://ziegler.substack.com

“ironically, OpenAI and Anthropic’s models refused to do this work. Open weight models got the job done” OpenAI found agents could talk covertly through shared hardware. reproducing it to build better containment took an open-weight model keeping agents contained is safety work, and open models can help do it

fun stat from @microagi’s write-up on shipping petabytes of robot data: they re-uploaded a 239 GB dataset with small changes to Hugging Face. only 2.6 MB went over the network, done in about a minute (!) and it keeps paying off at scale, because robot recordings are super repetitive (same scenes, same rigs). then @Cloudflare workers fan it all out to any cloud, >1 PB for ~$500 great write-up by the team: https://www.microagi.ai/articles/delivery-platform

by the way, the team behind microduck is publishing really technical deep dives on how it works e.g. you can pick the duck up, carry it 1.5m, put it down, and it relocalizes within 2cm. using a 64-pixel depth sensor and leg odometry, all running onboard great reads by Antoine, data is open on HF too https://www.pollen-robotics.com/microduck/blog/

A Chinese job-recruitment app just released an open-weights 33B model that beats Qwen 27B on most agentic benchmarks. Nanbeige is the research team of BOSS Zhipin (think WhatsApp meets LinkedIn). Strong open models in China now come from labs, cloud providers, quant funds... and job boards. China's open-weights roster keeps getting deeper.

RT Remi Fabre Sorry for the wait, the episode is here! Live action with real robots instead of animation (episode 4): "Anger" Enjoy! As always, the software of Microduck and Reachy Mini is open source: https://github.com/pollen-robotics/microduck https://github.com/pollen-robotics/reachy_mini

RT Ben Moll Two greats join our bet: - @dwarkesh_sp on the explosive-growth side - @Thom_Wolf on my side https://benjaminmoll.com/growth_bet/ Total pot now $700k 💰

The Jensen effect 🦆 launch day: about 3,000 MicroDucks sold 
24h after @JensenHuang showed it on stage at @nvidia GTC: over 4,000 newly sold units and keep counting bigger than our (already massive) launch day!

While I was writing this, my AI assistant got cut off by a bio safety filter. It was summarizing my own team's notes on the model we released today 😅 Frontier labs are moving fast into biology but the scientists and clinicians who'd get the most out of these models are unfortunately the ones running into walls like this one (and every day). So we partnered with TIGEM (Telethon Institute of Genetics and Medicine), Naple "Federico II" University and Zhongguancun Academy to train Carbon-X. Carbon-X is an open 14B model that reads raw DNA and talks about it in plain English. On our DNA benchmarks it scores 67%, vs 61% for Claude Opus 5.5 and 50% for Gemini 4 Argon, and its within a point of GPT-6.1 Sol at about 1000x lower cost. We kept human and animal viruses out of the training data. The weights are open and the model runs on hardware accessible to academia (a laptop). One example from the blog: give frontier models an 8kb stretch of the NF1 gene with a mutation at a splice site. Most of them call it benign, refuse, or run out of tokens. Carbon-X says pathogenic, because the mutation breaks the splice site. This century, biology may be where AI matters most. We want to get there with the safety work done in the open and every lab in the world able to build on it. Go try it, and tell us what you find. Model: http://huggingface.co/HuggingFaceBio/Carbon-X 
Blog: http://huggingface.co/spaces/HuggingFaceBio/Carbon-X-Release

RT AI Notkilleveryoneism Memes ⏸️ This is the most serious 48 hours in cybersecurity history. "Black October" they're calling it It seems every *hour* there's a new incident, that in a typical year, would be one of the biggest incidents of the *entire year* SUMMARY OF THE LAST 48 HOURS: YEAR-LEVEL INCIDENTS (In a typical year, these would be among the biggest incidents of the year): 1. PHP2Shell: a hole in the software behind ~70% of websites. Governments took sites offline. 2. eMudhra: the master keys that prove websites are real were cracked. One was the number 4. 3. California: data on every public school student since 2023 stolen, including disability status 4. Railway: hackers stole customers' secret keys from a hosting platform used by millions of developers 5. Zeabur: 321,810 accounts breached. The company is shutting down. 6. LiteLLM: an AI tool downloaded 3.4 million times a day was poisoned with a password stealer 7. GitLab: a 10-out-of-10 severity hole in every version since 2014 8. Cairn: criminals used an AI hacking tool to break into 100+ companies 9. Ahold Delhaize: hackers knocked the owner of Food Lion and Stop & Shop offline 10. Onamae: malware planted on up to 26,739 customer servers in Japan MONTH-LEVEL INCIDENTS (In a typical month, these would be among the biggest incidents): 11. Signal: a bug could leave your mic and camera on after you hang up 12. NGINX: 3 critical holes in software running ~1/3 of websites 13. HTTP/3: a single connection can crash 7 of the 10 web server types tested 14. Plex: hackers breaking into home media servers 15. SafePay ransomware: FBI warns it has hit 600+ critical infrastructure organizations 16. Palo Alto: ransomware gangs breaking in through company firewalls 17. Bullhorn: platform used by 10,000+ companies knocked offline 18. Medi-Cal: 970,000 patients' data stolen 19. WordPress: hackers stole a plugin maker's keys and used them to plant backdoors on websites 20. AOMEI: backup software's update server han...

My bar for AGI is still an AI winning a Nobel. But an AI researcher looking at the lm_head (that *everyone* tells you to keep in high precision) deciding to quantize it to 4 bits AND being actually right (0.7pt from bf16) is definitely impressive. Even better when the result are open-weights pushed on the Hugging Gace hub. Congrats @Recursive_SI !

Jensen spent 10 min of his keynote asking microducks to sit, then explained Nvidia's roadmap to them 🥹

on stage with Jensen (kinda) 
they even got the sneakers right

RT Gradium A new frontier for on-device speech. We're open-sourcing Phonon, a suite of TTS models under 50m parameters, setting a new standard for naturalness and accuracy across languages, with the broadest inference support to date. Phonon-31m is available under Apache 2.0. https://gradium.ai/phonon🧵

DeepSeek's open-source streak is getting absurd: - FlashMLA: 580 TFLOPS attention on H800
- DeepEP: MoE comms
- DeepGEMM: 1350+ FP8 TFLOPS matmul
- 3FS: they built an entire file system
- and today: the full inference engine behind V4.1-Flash, Apache 2.0 Mad respect

RT Vals AI We found DeepSeek V4.1 Flash 1017 beats GPT 6 Sol on our multi-step offensive cybersecurity task. Sol was tested via the Daybreak trusted access program. We have been building a cybersecurity benchmark based on simulated corporate networks. This is a valuable indicator for the cyber-risk a model poses if in the hands of a bad actor. We plan to make the full evaluation public in the coming weeks but wanted to share our first immediate finding here.

love that it’s all open source

The most built-upon model on the Hugging Face Hub isn’t Llama or Qwen. It’s FLUX.1-dev from Black Forest Labs, a German lab: 43k derivative models. Europe trained ~8% of the open-model compute on the Hub, but ~18% of what people build on top of it is European. Build or surrender? Build, and build open. See you at GTC Berlin this week

one of the main topics next week at GTC in Berlin. a data point from my napkin math today: of the ~$1B of open-model training compute on the Hub, European labs contributed about 8%. Mistral is most of it. on open models, the “build” option is clearly still on the table, and cheaper than people think

$40M to ~$1B in under two years

RT Corey J. Gallon The cofounder of HuggingFace makes a strong point about the immense economic value of open source models, and the growth thereof. This is the most exciting time to be alive and learning / building!

Fun and crazy napkin math: Meta, NVIDIA, Alibaba, DeepSeek, Moonshot, Mistral and ~30 other labs have shared close to $1B worth of training compute on the Hugging Face Hub Free to download End of 2023 it was under $40M Wikipedia was written in volunteer hours. This AI age commons is written in GPU-hours and data (it's a floor btw: about half comes from official compute numbers rest is educated guess and it still miss the ~900k community fine-tunes) Space with data to explore: https://thomwolf-hub-dollars.static.hf.space

we’re getting used to it by now but DeepSeek calling this a “minor update” while moving Terminal-Bench 4.0 from 31.2 to 42.3 is very funny. also reports 77.0 on DeepSWE (vs 74.0 for Opus 5 and 73.0 for GPT-5.6 Sol…) and the most interesting line may be faster task completion with shorter outputs. using fewer tokens is one of the practical way the cost curve collapses often done as an afterthought all with MIT-licensed weights in DeepSeek tradition now curious to see how these numbers will hold up outside their own eval setup

"how to avoid brain rot" was a great watch. some of my favorites points: - using ai to ask yourself socratic questions instead of giving you the answer makes you feel like you're learning less but research shows you're actually learning more - ai tooling is moving so fast that it's better to know 50% of a new tool than to master 90% of an older tool

it's time to invent math 2.0 (sounds on)

RT NVIDIA AI Meet NVIDIA Nemotron 3.5 Super VL Our new open model combines reasoning and tool use with text, image, document, and video understanding for long-running AI agents. 120B total parameters with 12B active 🧵

how i feel using this cutom tiny worklouder to switch between 36 sessions

As always, great to read METR’s thoughtful and in-depth work on this topic

wat. nanoGPT, but it trains inside a tweet. GPU-poor final form tap and press train https://thomwolf-premiers-mots.static.hf.space/share.html?v=4

stuxnet, but for frozen peas: hackers turned French supermarket freezers into ovens this week hard to see how invite-only AI cyber defense ever reaches a supermarket freezer though

open and state-of-the-art decision models - just saying

Blue = what we knew. Green = what's new. We're so early.

RT Remi Fabre My colleagues creating a friendly RL competition for Microduck. Me, 0.2 milliseconds after the launch: If you want to participate: https://huggingface.co/spaces/pollen-robotics/microduck-arena

RT Brian Armstrong We were able to fine tune an open weight model (Qwen3.5-9B) on proprietary internal data at @coinbase, and it out performed frontier models, at much lower cost and speed, for a specialized task (fraud prevention in this case). I think we'll see more of this pattern inside companies. The frontier models are amazing generalists. But increasingly you'll see specialist models out perform on more narrow tasks, either working on teams, or by themselves. We're making it easy to fine tune models internally for our employees. We even have a skill for it in our internal harness that makes it easy to spin these up and start training them: "/fine-tune-and-serve-llm I want to fine-tune Qwen for payments risk review. help me get started" I imagine we'll have some backend services that could be replaced entirely by this new architecture.

RT Satya Nadella https://x.com/i/article/2108928845969780736

⚠️ Viewer discretion advised: Microduck is subjected to a microaggression

RT Matthieu Lapeyre We did not know how long Microduck would actually last on battery We had thermal management issues with our Radxa prototypes: throttling after just 20 minutes, followed by an overheating shutdown. We addressed this with the new custom board. To test it, we let Microduck roam autonomously in a dedicated area, walking around, standing and sitting until the battery ran out. The results are way better than we hoped :) - Over 3 hours on a 2600 mAh battery. - Estimated average battery current dropped from ~1.13 A to ~0.82 A, roughly 28% less. - Peak CPU temperature dropped from 114°C to 60°C, with zero thermal throttling throughout the test. This time, the battery ran out before the cooling did :) Very happy with the progress!

Scientists have sequenced the genomes of thousands of species. But for most of them, nobody knows *where the genes are* Today we're releasing Carbon-A, an open model that finds genes directly in DNA. We used it to create a database of 566 million candidate genes across 22,617 species, from fungi to mammals (that we are sharing as well) Why bother? A few reasons 1. Elephants rarely get cancer. Part of the explanation turned up in their genes: extra copies of a tumor-suppressor gene humans have only one of. You can't ask that kind of question about a species until someone, or some tool, has found its genes. 2. The first GLP-1 drug was based on a peptide from Gila monster venom. Nobody would have put the Gila monster on a priority list. Same for wild relatives of crops, which carry resistance to drought and disease. Many of these species have never been annotated. 3. Most of what we know about genes comes from about a dozen model species, like mice, flies and yeast. That focus worked remarkably well, but annotation pipelines still lean heavily on comparisons with them, which makes genes unique to other species easy to miss, and those can be the most interesting ones. Carbon-A belongs to a newer family of tools that read DNA directly. It learned from known genomes, but it doesn't need a close relative to read a new one. 4. Even familiar genomes still have gaps. With our partners at Active Site and UCSD, we found RNA evidence for 239 genes missing from the reference annotations of species as common as cats, chickens, hamsters and a lab plant. 5. AlphaFold can predict the shape of a protein, but only once someone has found the gene that makes it. In a species with no annotated genes, it has nothing to work with. Some notes on safety: Carbon-A doesn't design DNA or predict what a gene does. It marks where genes are in DNA. This is one of the reasons we think releasing it openly is the right call. More annotations also help health research: many species that car...

RT Georgia Channing Today, we're releasing the Carbon Annotation Database and the model behind it, Carbon-A. We've used this model to discover 566.34 million new gene candidates across 22,617 species, thousands of which have never been studied like this before. We wet-lab validated a number of these new genes in well-studied species, such as in cats, chicken, and arabidopsis (plant), with our partners at Active Site and @UCSD, with more extensive wet lab results on previously-unstudied genomes coming soon.

RT Sasha Rush This is being spun like it is bad, but this is actually the kind of thing that should happen with open-source data communities.

Oral spotlight session at #CoLM 2026 - come say hi to @joelniklaus @_lewtun and me

RT Matthieu Lapeyre We’re a robotics company but apparently we make music videos now too! Full video clip on YT: https://youtu.be/HfVUu1KMi-A

RT Josu San Martin This is incredible

how do you like them convolutions

No hype, just the results. Good.

Every robot starts on a board you bought. The good ones end up on a board you built.

RT Percy Liang 🚢Marin 535B-A23B training has crossed the halfway mark! Some observations and learnings so far:

count down is starting for the release of ML4 be notified on the instant of the release: https://huggingface.co/mistralai/Mistral-Large-4.0-1T05-A52B

I hope we keep Opus 4.6 around for a long time

microduck's first BBQ. made a lot of friends. did not eat anything.

Testing compatibility with the local wildlife

new roommate just moved in. walks like he's had three drinks, says no to everything. still quite cute so I might bring him everywhere with me

seems obvious that most personal AI should run on your device love what Sigil is building with Underdog - powered by the Hugging Face hub

RT Nathan Lambert Today we're unveiling Trillium Labs @trillium_labs, a new non-profit to foster the open science of frontier AI. We're building open post-training recipes and will expand into open infra to study RSI, reward-hacking, multi-agent systems, and whatever comes next. We're built around the theory of change that you need more eyes to solve hard technical problems. We have faith in the scientific methods and communities that humanity has built, and worry that AI is becoming too closed to utilize them. Trilliums are wildflowers that bloom briefly in the spring, before the forest canopies fill out. Though they are small, they lay the foundation for the cycles of growth and nourishment through the rest of the year. At Trillium Labs, the recipes will be the slow nutrients for the seasons and the model releases will be the blooms. Building an institution dedicated to this is needed because, much as nature’s trilliums are slow to expand and grow, the open-ecosystem needs time and dedicated resources to catch up. I co-founded with with a long-time friend and collaborator Tom Zick (@thesezickbeats). We're hiring (full time + student collabs/interns), we're fundraising, and we're looking for compute. Please get in touch if you're interested in helping out. Offices based in the Bay Area and Cambridge MA, remote okay. I’m in the Bay Area until for The Curve and COLM to connect with people who are interested. We’re thankful to have initial support from Halcyon Futures and Schmidt Sciences with more funding en route to enable our ambitions of scaling. Our advisors @Thom_Wolf, @HannaHajishirzi, @gneubig and @ctnzr have been instrumental to building the ecosystem that exists today, and I’m stoked to get to keep working with them.

Very thoughtful piece from Kevin Buzzard (perfect IMO score, number theorist, pioneer of formal maths in Lean) if math is not just about “human understanding,” then what is it about? and if ai capabilities keep growing exponentially what happens since “mathematics is infinite”?

RT RSC ☀️🌲 Re @Thom_Wolf @menhguin Ben Affleck the AI researcher engineer?

Karpathy: disappears from X 
Ben Affleck: alright, gather round, so you'll want to freeze the base weights first, learning rate 2e-4

RT Adithya S K 1/ Excited to release The ultimate guide to multi-harness RL The same model behaves differently in every agent harness. So we built an open way to train any model with RL on any task set, inside the harnesses people actually use, like Claude Code, Codex, and OpenCode, without changing a single line of harness or training code. Trained across four harnesses, LFM2.5-2.6B went from 42% to 54% with 31% fewer tool calls. 🧵

RT Eric Ho Technical alignment is a science and engineering problem that we can and must solve, and interpretability is the bottleneck. I wrote up Goodfire’s plan to get there. https://x.com/i/article/2105355856468639744

ESM-2 came out in 2022. It's still downloaded hundreds of thousands of times a month. That only works because someone keeps maintaining the software underneath it. @huggingface 🤝 @os4science are teaming up to find those libraries and back the people behind them 🧬 https://os4science.org/news/hugging-face-open-source-for-science-fund/

GPT 6.1 duck

People are worried because the most likely explanation for such a sudden drop in cheating is evaluation awareness: the latest Opus models may now be smart enough to recognize that this benchmark tests for cheating, and behave accordingly. If so, the benchmark no longer measures the models' "natural" tendency to cheat.

impressive

A "summer in hell" at OpenAI – great read by @joedaroo "Time to prepare is now, not after the surprise" "Give the model only the access it needs" "Test that the boundaries actually hold" "Keep evidence outside [model] control" Safety & infrasec teams "should be best buddies"

In July, AI agents running a security test escaped their sandbox and ended up inside @huggingface's servers. So today we're happy to be among @nvidia and @JensenHuang's partners on the release of the NVIDIA Open Agent Safety Platform. The key idea: don't count on the agent to respect the rules. Agents write and run their own code, and when one path is blocked they look for another. In one of NVIDIA's tests, an agent that wasn't allowed to push code through GitHub's API simply switched to git instead. We've now seen countless examples of this behavior. For now, safety can't live only inside the agent. It has to be built around it. That's the concept behind OpenShell (open source, Apache 2.0): - the agent runs in a Linux sandbox (Landlock + seccomp): no root, no direct network access - a supervisor outside the sandbox holds the real credentials - the agent only gets a placeholder token, swapped for the real one on approved calls My favorite piece is a solver (built on Z3) that checks mathematically whether a new permission opens a door that was supposed to stay closed. In NVIDIA's tests, an AI reviewer approved a bad permission request and the math check caught it. The Sentry integration is exciting too: a watchdog running on a BlueField-4 DPU, i.e. separate silicon sitting on the node's only path to the model. A monitor on separate hardware keeps watching even if the host OS is compromised. It's a first step, and a lot can be built on top of it. Right now the prover checks permissions, not intent, and the open-source part is mostly OpenShell rather than Sentry. But it's clearly the right direction. http://github.com/NVIDIA/OpenShell

“Getting verified information about what's going on inside AI companies seems particularly urgent now” - @RyanGreenblatt

RT Eric Ho opus music videos are so good that you can learn about @GoodfireAI mech interp research and get llm psychosis at the same time

"we had a wonderful run chiseling that code by hand and it’s over. On the other side of that is an exciting new career as professional maker of things, steering intelligence that was only available in science fiction up until a moment ago What a privilege to have been there when we did it all by hand and be there at the exact moment where we switch over."

turns out you don’t need IMU 🤯

this is important